Spotify Music Playlist RecommenderClaire Medina, Ferdie Taruc, Psalm Masaya

Team

Context● Streaming services revolutionized accessibility of music

Given user preferences, how can we create a playlist that can encompass the interests of different people?

Research Question:

Motivation● Creation of shared playlists with others● Catering to different interests ● Current Projects: Recommending Songs based on mood, etc.● This Project: Appealing to Different Users

Definition of “Playlists”● Unit composed of songs

in some ordering

● Subjective

● Definition: Group of songs that fit a certain theme or genre

Limitations to this Definition● Why this definition?

○ Difficulty in Dimension Reduction

● Can no longer capture:

○ Context of the playlist

● Content -based filtering

Music Recommendation Systems

● Collaborative Filtering:○ Exploits interactions between users and items○ Identifies similarities between users

● Content Based Filtering:○ Intrinsic Audio Features○ Textual Metadata

● We will use the latter method of filtering for our model.

Introduction to Approachestake .json files convert dataset into dataframe

find optimal k given data, cluster using k -meansgiven song input, create a playlist of similar songs

Spotify Web API● Spotify’s Web API

● Spotipy

● Other interesting audio features it provides:

○ Danceability, energy, instrumentalness, liveness, loudness, speechiness, valence (positiveness), tempo (BPM)

○ Explicitness, Release Date of Song

More information about audio features: https://deve loper.spotify.com /docum enta tion /web-ap i/re fe rence /tracks/ge t-aud io-fea tu res/

Spotify Million Playlist Dataset

● Enable research for music recommendation systems in 2018.

● Captures 1 million playlists of over 2 million unique tracks.

● Re-released September 2020 by AICrowd

Dataset source: https:/ /www.a icrowd.com /cha llenges/spotify-m illion-p laylist-da tase t-cha llenge

Dataset Statistics● All playlists were generated by users from the United States:

○ Gender■ 45% Male | 54% Female | 0.5% Unspecified | 0.5% Nonbinary

○ Age■ 13-17: 10%■ 18-24: 43%■ 25-34: 31%■ 35-44: 9%■ 45-54: 4%■ 55+: 3%

Dataset: Playlist Criterion● Playlists that meet the criteria are selected at random:

○ Contains at least 5 tracks and no more than 250 tracks○ Contains at least 3 unique artists and at least 2 unique albums○ Was created after January 1, 2010 and before December 1, 2017○ Does not have an offensive title○ Etc.

● Other modifications made to some playlists

Dataset Formatting● Approx. 33 GB

● Stored as .json files

○ Approx. 1000 slices

○ 1000 playlists within each slice

● Huge Database

Example of .json

● Playlist information provided:○ Name of playlist ○ PID (unique identifier)○ Collaborative○ Last Modified○ Etc.

● “URI” (song and artist identifier) used to query specific information from Spotify’s API.

Dataframe Example

Dataframe Example (2)

Adding Artist Genres ● Approximately 5,071 distinct genres to

describe artists in Spotify

● Echonest applies ML on audio features to classify genres

● Classify genre of song based on genre of artist (has limitations!)

Adding Artist Genres (everynoise.com)

atmospheric bouncy/upbeat

Mechanical instrumentation (techno music)

Organic instrumentation (classical)

T-SNE VisualizationOrganic Instrumental/ More Classical

Upbeat and Radio Edited Songs-Britney Spears-Drake

Slower Music like Jazz or Christian Music

Pop Songs with a lot of Mechanic Instrumental

The Playlist Recommendation System

K-Means Clustering

Elbow Method

Building a Playlist● Input = Song Name, Artist Name, Spotify DataFrame● Randomized selection of 30000 rows from csv due to hardware limitations

Building a Playlist● Train the KMeans model on the given DataFrame● Append cluster labels to the DataFrame for each row using kmeans.labels_

Building a PlaylistPreprocessing for the input.

Predict cluster label for input song.

Building a Playlist● Filter out dataframe by cluster label● Calculate euclidean distances and sort based on minimum distance● Pick top n songs (n = 2 for example)

Building a Playlist: n -songs Input

Inputs:

● List of Lists → [ [Song Name, Artist Name], [Song Name, Artist Name],... ]● Spotify DataFrame● Number of Songs to Recommend / Song

The Example

Output

Output

Output

Limitations of Dataset● Only used 4 slices (500MB) of .json file s

● Lim ited to on ly a sm all am ount of songs

● Playlist crite rion a lso crea tes a b iased sam ple .

● Unable to represen t a ll genres

● Sm alle r in com parison to la rge r recom m endation system s like :

Conclusion

● Real-World Implications○ Ability to create a playlist off of different user preferences

Next Steps: Ready for the Real-World?

● Ready for the Real -World?○ Scale Up○ Playlist metadata○ Ensemble of Collaborative Filtering with Content Based techniques

Thank You!

Questions?

Spotify Music Playlist RecommenderTeamContextGiven user preferences, how can we create a playlist that can encompass the interests of different people?MotivationDefinition of “Playlists”Limitations to this DefinitionMusic Recommendation SystemsIntroduction to ApproachesSpotify Web APISpotify Million Playlist DatasetDataset StatisticsDataset: Playlist CriterionDataset FormattingExample of .jsonDataframe ExampleDataframe Example (2)Adding Artist Genres Adding Artist Genres (everynoise.com)T-SNE VisualizationThe Playlist Recommendation SystemK-Means ClusteringBuilding a PlaylistBuilding a PlaylistBuilding a PlaylistBuilding a PlaylistBuilding a Playlist: n-songs InputThe ExampleOutputOutputOutputLimitations of DatasetConclusionNext Steps: Ready for the Real-World?Thank You!Questions?